Preserving HugeTLB Memory During Live Kernel Updates
At the 2026 Linux Storage, Filesystem, Memory Management, and BPF Summit, Pratyush Yadav led a memory-management-track session on a critical feature: preserving hugetlbfs (HugeTLB) memory across live kernel updates. Live updates (using kexec handover) allow a new kernel to replace the running one without a full reboot, but large pages from HugeTLB were previously lost—forcing applications to reallocate. Yadav's proposal aims to keep these huge pages intact, improving performance and reliability. Below, we answer key questions about this development.
What exactly is HugeTLB memory and why is it important?
HugeTLB (Transparent Huge Pages' big brother) provides hugetlbfs—a filesystem that allocates large, physically contiguous memory pages (e.g., 2MB or 1GB) for applications like databases, scientific simulations, and virtual machines. These huge pages reduce page table overhead and improve TLB (Translation Lookaside Buffer) hit rates, drastically speeding up memory-intensive workloads. Without them, performance can suffer significantly—imagine a database caching multiple gigabytes: HugeTLB ensures minimal CPU cycles wasted on page table walks.
What is a live kernel update (kexec handover)?
A live update (or kexec handover) lets the kernel swap itself for a new version without a full reboot. It uses kexec to load the new kernel into memory, then performs a “warm” handover: critical state is saved, hardware reinitialized, and userspace processes continue. This is vital for servers that cannot afford downtime. However, until now, hugetlbfs memory was not preserved—applications had to release and reallocate huge pages, negating their benefits temporarily.
What problem did Pratyush Yadav's session address?
Yadav tackled the loss of HugeTLB memory during live updates. When the old kernel hands over to the new one, hugetlbfs pages were freed and then re-requested by applications after boot. This caused a “performance dip”—the very advantage of huge pages was lost. Prescott Preservation intends to keep those pages mapped across the handover, so applications see no disruption. The session at LSFMM2026 focused on design challenges and a proposed kernel infrastructure to achieve that.
How does the proposed preservation mechanism work?
The mechanism extends the kexec handover path to treat hugetlbfs pages as a special resource. Instead of freeing them, the old kernel passes metadata (physical addresses, page sizes, flags) to the new kernel via a reserved memory area. The new kernel then re-creates the hugetlbfs pool from that metadata, mapping the same physical pages into the same virtual address space. Key steps include:
- Freezing HugeTLB allocations before handover
- Serializing page table entries and pool state
- Reconstituting the pool in the new kernel before resuming userspace
This ensures zero loss of huge pages.
What challenges did Yadav highlight?
Three major challenges were discussed:
- Page table coherence – The old kernel’s page tables reference huge pages; the new kernel must rebuild them identically without conflict.
- Memory accounting – Both kernels must agree on which pages are hugetlbfs to avoid double-allocation or leakage.
- Driver interactions – Some subsystems (e.g., RDMA) pin huge pages; re-pinning after the handover must be seamless.
Yadav proposed a reserved metadata region to handle coherence, and hinted at a new kernel interface for driver notifications.
What are the expected benefits for users?
For administrators and application developers, the feature means downtime-free kernel updates without performance impact. A database upgrading the kernel can keep its 1GB HugeTLB pool active—no need to warm up caches or reallocate massive pages. Scientific clusters running MPI jobs will also see smoother transitions. Combined with the live update orchestrator, it enables true non-disruptive maintenance for large-scale deployments.
Where can I learn more or contribute?
Keep an eye on the Linux kernel memory-management mailing list for patches from Pratyush Yadav. The session at LSFMM2026 is recorded (attendees only). For now, reviewing the LWN summary and joining the discussion on HugeTLB is recommended. Contributors familiar with kexec, memory management, or hugetlbfs are especially welcome.